1Fast overview
Wright's Law says that for every doubling of cumulative production of something — total units ever made, not units per year — the cost of producing one more unit falls by a roughly consistent percentage. Make the second one, and it is a little cheaper than the first. Make the two-thousandth, and it is a lot cheaper than the thousandth, by the same proportion each time.
The idea traces back to 1936, when engineer Theodore Wright studied aircraft factories and found labor costs per plane kept dropping on a predictable curve as total output grew. A few decades later, the Boston Consulting Group generalized the same pattern — renamed the "experience curve" — to cover a business's total costs, not just labor, across many industries: ships, semiconductors, solar panels, batteries.
The mechanism is learning, not magic: workers get faster, engineers redesign for cheaper manufacturing, suppliers specialize, waste gets found and removed. None of that happens by calendar time alone. It happens by doing the thing, repeatedly, and paying attention.
Cost per unit falls by a fixed percentage — commonly 10–25%, depending on the industry — every time cumulative production doubles.
Cost(N) = Cost(1) × Nlog₂(r)
where N is cumulative units made and r is the "progress ratio" (e.g. 0.85 for a classic 15% learning rate).
2Mental Model Compass
3Learn it, step by step
It is not the price you charge that falls on its own — it is the underlying cost to produce one more unit. That cost falls because of four ordinary, unglamorous mechanisms: workers and teams get faster at the specific tasks (learning by doing), engineers redesign the product and the process once they see where the waste is, suppliers and tooling specialize around the now-proven design, and organizations find and eliminate small inefficiencies they didn't notice the first few times. None of this requires a breakthrough. It requires repetition plus attention.
Click to reveal: Why does this track cumulative output instead of calendar time?
Because the learning comes from the doing, not the waiting. A factory that makes 10 units a year for ten years has the same cumulative experience — and should sit at roughly the same point on the curve — as one that makes 100 units in one year. Time elapsed doesn't teach anyone anything; repetitions do.
A common mistake is to think of the discount as "costs fall X% per year." They don't — they fall a roughly fixed percentage per doubling of total units ever made. Early on, doublings happen fast (going from 10 to 20 units is one doubling), so costs fall quickly in the beginning. Later, doublings take far longer (going from 10 million to 20 million units is also just one doubling), so the same percentage drop is spread over a much longer stretch of calendar time. This is exactly why early-stage technologies look like they are improving at a blistering pace, and mature ones look like they've plateaued — even when the same underlying law is operating the whole time.
Wright's Law is an empirical regularity, observed across many industries with remarkable consistency — but it is not guaranteed. It can stall (a design hits a genuine physical limit), reverse (a key input becomes scarcer or more expensive even as production grows), or accelerate (a real scientific breakthrough stacks on top of ordinary learning). Treat the curve as a strong prior to test against real data for your specific product or industry, not as a formula you can extrapolate blindly ten years out.
The practical move is almost always forward-looking: if you believe a cost will keep falling on its experience curve, the question becomes whether to invest now, ahead of demand, to buy your way onto the curve faster than a slower-moving competitor — effectively trading cash today for a lower cost position tomorrow. This is a real strategic bet, not a free lunch: you are spending real money on the assumption that the curve holds and that volume actually materializes.
4Examples across different worlds
Aircraft manufacturing, 1936. Theodore Wright's original study found that for every doubling of total airframes built, direct labor hours per airframe fell by a consistent percentage. The military used this to negotiate forward pricing on aircraft contracts before the next batch was even built, because the curve was that reliable.
Semiconductors. Chip fabrication costs per transistor have historically fallen on a steep experience curve as cumulative wafer production has grown, layered on top of (and often confused with) the separate density gains of Moore's Law. Fabs invest billions ahead of demand specifically to capture this curve before a competitor does.
Solar photovoltaic panels. The cost per watt of solar panels has fallen roughly 20–24% for every doubling of cumulative global installed capacity over several decades — one of the most-cited real-world experience curves, and a central input to climate and energy policy modeling about how fast renewables become cost-competitive without subsidy.
Lithium-ion batteries. Battery pack costs have followed a comparable curve as cumulative production (driven heavily by electric vehicles and grid storage) has scaled, which is why government and industry forecasts for EV price parity lean so heavily on projected cumulative volume, not just calendar years.
Learning a skill. Your hundredth cold call, your hundredth client presentation, your hundredth home-cooked version of a recipe — each one takes noticeably less effort and produces noticeably better results than your first few, following the same doubling logic on a personal scale. Ten repetitions in you are nowhere near your most-improved point; eighty, one hundred and sixty, and so on are where the curve keeps paying out.
Career choice. Taking a role with high repetition and fast feedback (many client pitches, many code reviews, many negotiations) puts you further along your own experience curve faster than a role with the same total years but far fewer actual repetitions of the specific skill.
AI inference pricing. The cost to run a given level of AI capability has fallen sharply as cumulative deployed usage and hardware production have scaled — a pattern widely discussed by AI research groups tracking historical compute-cost trends. Forecasting next year's API pricing by projecting last year's percentage drop in a straight line, instead of thinking in terms of cumulative volume and hardware generations, is a common and avoidable error.
Evaluating an AI vendor's roadmap claim. When a vendor promises a cost curve will keep dropping at historical rates, the useful question is "what cumulative volume is that projection assuming, and is anything likely to interrupt it" (a chip shortage, an energy constraint, a genuine architecture plateau) — not simply trusting the trend line to extend itself.
5Worked application: pricing a new manufactured product
Scenario: your company is about to launch a new hardware product. Leadership is debating whether to price it near cost at launch (thin or negative margin) or price for healthy margin from day one.
- Estimate today's unit cost and progress ratio. Your first production run costs $400 per unit to build. Based on comparable products in your industry, costs have historically fallen about 15% per doubling of cumulative units (a progress ratio of 0.85).
- Project the curve forward. At 0.85 per doubling, by the time you've built 8× your current cumulative volume (three doublings), your cost per unit should fall toward roughly $400 × 0.85³ ≈ $246 — before any other changes.
- Decide whether volume is the bottleneck. If your main obstacle to reaching that cumulative volume is weak demand at your current price, then pricing near cost now — to sell more units sooner — directly accelerates your own trip down the curve. This is the strategic logic behind many "sell at a loss to build volume" launches.
- Check who else is racing the same curve. If a well-funded competitor is pricing aggressively to build cumulative volume faster than you, delaying your own low-price push may let them reach a durable cost advantage first — one that is very hard to close later, because they will simply have made more units than you ever can catch up to at the same rate.
- Stress-test the assumption. Ask what would break the curve for your specific product: a scarce raw material, a regulatory cap on production, a technology with limited further room to improve. If any of these apply, discount your projected future cost accordingly rather than trusting the historical percentage blindly.
- Make the pricing call explicitly. Price for volume only if you have confirmed (a) the progress ratio is grounded in real comparable data, not optimism, and (b) you can actually finance the gap between today's cost and today's price long enough to reach the volume where the curve pays you back.
6Common mistakes
Confusing it with Economies of Scale. Wright's Law is about cumulative history (total units ever made); Economies of Scale is about present-day size (a bigger factory is cheaper right now, with zero history). A brand-new giant factory gets a scale discount on day one but has not yet earned a single unit of experience discount.
Projecting the percentage drop per year instead of per doubling. Because doublings get further apart in calendar time as volume grows, a straight extrapolation of last year's percentage drop will systematically overstate near-term progress and understate how long later gains actually take.
Assuming the curve applies to any repeated activity regardless of whether anyone is actually learning from it. Mindless repetition without feedback, redesign, or attention to waste does not reliably produce the discount — the mechanism is learning, not mere elapsed repetition.
Treating the historical progress ratio as a permanent constant. Progress ratios drift — curves can steepen when a new technique arrives, or flatten when a technology approaches a physical limit. A ratio measured from the last decade is a starting estimate, not a guarantee for the next one.
7When not to use it
You have very few historical data points. Fitting a progress ratio to three or four observations and then projecting a decade forward is overconfident curve-fitting dressed up as economic law.
The product is fundamentally changing, not just scaling. If each new unit is a materially different design (one-off custom projects, bespoke consulting engagements), there is no stable "cumulative experience" to measure, because you are not really repeating the same task.
A hard physical or regulatory ceiling is already visible. If a known limit (a material's theoretical minimum cost, a legal cap on output) sits close to today's position on the curve, further doublings cannot deliver the same proportional gains no matter how much volume accumulates.
You are being sold a growth story. "Costs always fall with scale, so trust our long-term projections" is a convenient claim for anyone raising money or defending a valuation — reasonable to ask for the actual historical data behind the specific progress ratio being assumed.
8Related models in this series
9Try it yourself
Answer three quick questions and get a ready-to-use prompt for thinking through your own situation with this model.
10Save as a skill
A portable version of this model you can hand to an AI assistant so it applies Wright's Law whenever it's relevant to what you're working on.